Research Summary

Abstract:

Digital environments increasingly shape political opinion through recommender systems and large language models (LLMs). This study proposes a hybrid mathematical model in which polarization emerges from interactions among social influence, bounded confidence, biased assimilation, algorithmic personalization, and generative responses. The model distinguishes structural from effective influence networks and treats LLM output as a stochastic component requiring independent calibration. It supports comparative simulation, sensitivity analysis, and validation. The central argument is that personalization is not inherently polarizing, but may reinforce polarization when combined with selective exposure, narrow confidence bounds, and biased assimilation. LLM agents are therefore modeled as generative components rather than substitutes for human citizens.